CCLDNet

CCLDNet enhances medical image segmentation accuracy by combining Conditional-Synergistic Convolution (CSConv), a Lesion Decoupling Strategy (LDS), and a transformer backbone to improve polyp and skin lesion delineation for diagnostic and monitoring applications.

Key Features:

• Conditional-Synergistic Convolution (CSConv): Dynamically generates lesion-specific convolution kernels to adaptively model the unique characteristics of individual lesions.
• Lesion Decoupling Strategy (LDS): Decomposes the lesion segmentation map into lesion center and lesion boundary soft labels to simplify boundary delineation.
• Transformer Network Backbone: Replaces fixed CNN structure with a transformer backbone to enable global dynamic modeling across the network.

Scientific Applications:

• Polyp Segmentation: Achieved an 89.22% dice score on the EndoScene benchmark.
• Skin Lesion Segmentation: Achieved a 91.15% dice score on the ISIC2018 dataset.

Methodology:

CSConv dynamically generates specialist convolution kernels per lesion, LDS splits segmentation into lesion center and boundary soft labels, and a transformer backbone provides global dynamic modeling.